One of the most prevalent complications of diabetes and a leading cause of preventable blindness in the world is DR. This paper examines the history of the DR detection methods, the transition of traditional image-processing algorithms to the state-of-the-art deep learning and ensemble models. This study employed MESSIDOR dataset; it includes retinal fundus images with labels indicating DR severity. Images were enhanced and pre-processed, and extracting features based on deep learning was done using pre-trained convolutional neural networks (CNNs). In particular, the models that were used to extract the features included ResNet50, InceptionV3, DenseNet121, DenseNet169, VGG16, Xception, and a custom Deep Neural Network (DNN). Moreover, ensemble methods were designed through integrating CNN-based feature extractors with machine learning classifiers such as Random Forest, Support Vector Machine (SVM), XGBoost and LightGBM to improve the accuracy of classification and generalization. The results of the experiments showed that among the autonomous deep learning models, Xception achieved the highest performance, with a testing accuracy of 95%. EfficientNetB3 combined with XGBoost achieved the highest testing accuracy (86.83%). This study highlights the possibilities of deploying deep learning and ensemble AIs into DR screening devices, especially in telemedicine systems and mobile diagnostic systems. Such integration could significantly improve early detection rates, reduce the burden on healthcare providers, and make diabetic eye care more accessible in underserved regions.